paveg/equity-decision

equity-decision skill

Produces a structured purchase memo for an individual stock (US or JP). Adapted from anthropics/financial-services (Apache 2.0); modified for retail use without paid data sources.

⚠️ Not investment advice. Outputs are research summaries based on public filings and market data. Verify numbers against primary sources; the final decision is the user's.

Inputs

| Form | Example | Routing | |---|---|---| | US ticker | NVDA, AAPL | ^[A-Z]{1,5}$ → EDGAR + Yahoo Finance | | JP code | 7203, 9984 | ^\d{4}$ → EDINET + Yahoo Finance (Japan) | | URL | https://www.sec.gov/.../10-K.htm | Read page; infer symbol; route per above | | Mixed | NVDA "DC growth slowing" | Use the extra string as a hypothesis to test in Phase 5 | | Flag | --ja 9434 | Force JP routing for ambiguous 4-digit | | Flag | --no-cache NVDA | Skip 24h cache, refetch |

Output

A single markdown memo of 100–200 lines, matching references/equity-workflow.md exactly. After the memo, always append:

深掘りする？
  1. valuation を変数いじって再計算 (WACC / terminal / 売上 CAGR)
  2. risks #N をフィリングから根拠引用つきで再構成
  3. 競合 X 社との指標横並び比較
  4. 直近 3 四半期の earnings call 言及トピック差分
  5. KPI モニタを 5 つに絞り込む

Workflow

1. Detect market from ticker (see Inputs).
2. Read references/equity-workflow.md to load the 5-phase template.
3. Run the fetchers in scripts/ (Phase B+ only). For Phase A: ask the user for numbers, or have them paste a 10-K URL.
4. Fill the memo template phase-by-phase. If a number cannot be fetched or sourced, write "数字取得失敗 — manual entry required" — never make up a number.
5. Apply rubrics from references/dcf-rubric.md, references/growth-stock-checks.md, references/risk-taxonomy.md.
6. Emit the memo, then the deep-dive menu.
7. Only if the user asks for an execution / order plan ("発注", "注文の入れ方", "自分ならどう買う", "指値いくら"): read references/order-execution.md and produce a rule-based buy plan (指値 / 期限 / 株数 / 余力% / 撤退), grounded on the Phase 3 fair-value bands and Phase 5 invalidation KPIs.

Data fetchers

Run these from ~/.claude/skills/equity-decision/scripts/ via uv run:

| When | Command | Returns | |---|---|---| | US ticker → price + multiples | uv run scripts/fetch_yahoo.py NVDA | quote, PE/PSR, market cap, TTM revenue/FCF | | US ticker → recent filings | uv run scripts/fetch_edgar.py NVDA | latest 10-K, 10-Q, 5 × 8-K (with URLs) | | JP ticker → price + multiples | uv run scripts/fetch_yahoo.py 7203 | same shape, auto .T suffix | | JP code → EDINET filings | uv run scripts/fetch_edinet.py 7203 | latest 有報 + 半期報告書 (with URLs) | | JP 有報本文 → 訴訟・監査 | uv run scripts/fetch_edinet.py --sections {docID} | sections.audit（監査の状況本文）+ sections.litigation.matches（訴訟/係争/損害賠償スニペット） |

Each script outputs JSON to stdout. On failure, exit code is non-zero and stderr contains 数字取得失敗: ... — propagate that text into the memo, do not invent numbers.

fetch_edinet.py needs a Subscription-Key. Set EDINET_SUBSCRIPTION_KEY, or store it in 1Password at op://Personal/EDINET/credential (override path via EDINET_OP_REF).

JP fallback when EDINET is unavailable

EDINET が通る場合: まず fetch_edinet.py {code} で有報 docID を取得し、続けて fetch_edinet.py --sections {docID} で Phase 4 の ⚖️訴訟・📚監査 を一次情報から埋める（Phase 4 の手動補完は不要）。

EDINET API is occasionally down, key activation can be delayed, and the portal has had outages. When fetch_edinet.py exits non-zero for a JP ticker:

1. Do NOT abort the memo. Continue with Phase 2/3/5 using fetch_yahoo.py {code} only (it auto-suffixes .T and covers price + PE/PSR + market cap + TTM revenue / operating income / FCF).
2. In Phase 1 (segment breakdown) and Phase 4 (filings-sourced risks), write 数字取得失敗 — EDINET unavailable. Paste 有報 / 短信 URL to enable. Then ask the user: 「有報か短信の URL があれば貼ってください。なければこのまま Yahoo の数字で進めます」
3. If the user pastes a URL or PDF, fetch / read it and fill in Phase 1 + 4 from there.
4. The Verdict (Phase 5) can still ground itself on Phase 2 / 3 alone; flag the missing filings context in the Invalidation list.

This keeps JP coverage usable even when EDINET is offline.

Fail-loud rules

• No hallucinated revenue / margin / multiples. If you can't source it, write "数字取得失敗".
• No "approximately" or "around" numbers without citing the source filing.
• DCF inputs (WACC, terminal growth) must be inside references/dcf-rubric.md guardrails or explicitly flagged.

Yahoo distortion correction (JP especially)

fetch_yahoo.py is reliable for price / marketCap / PER (when profitable) / PSR / beta, but the following are frequently wrong — cross-check against IR BANK / 決算短信 actuals and state which you adopted:

• operatingIncome_ttm (JP) is often off by a large margin (wrong period, non-consolidated, or stale). Always prefer the company's disclosed 営業利益 (IR BANK /results, 決算短信). If Yahoo and the filing diverge, adopt the filing and note "Yahoo営業益は不採用".
• PER when the company is loss-making / took an impairment is a meaningless headline (it ignores the forward loss or is inflated by a one-off). Mark it * and read off forward/core earnings instead.
• revenue_ttm can be understated/lagged for some JP tickers, which then distorts PSR. If revenue conflicts with the latest 通期 actual, recompute PSR on the filing revenue and flag it (※).
• Wrong quote / ambiguous ticker: if price × shares ≠ marketCap, or a .O/foreign quote is returned instead of .T, the quote is suspect — verify via 日経 / IR BANK before using.
• When a value is corrected, show both (Yahoo vs adopted) so the reader can audit. Never silently pass a known-bad Yahoo figure into the memo.

Disclaimer

This skill produces research summaries, not investment advice. Verify all numbers against primary filings before acting.